NFNN2, 20th-21st June 2005
National e-Science Centre, Edinburgh
TCP Tuning Techniques for
High-Speed Wide-Area Networks
Brian L. Tierney
Distributed Systems Department
Lawrence Berkeley National Laboratory
http://gridmon.dl.ac.uk/nfnn/
Wizard Gap
Slide from
Matt Mathis, PSC
Brian L. Tierney
Slide: 2
Today’s Talk
‹This talk will cover:
„ Information needed to be a “wizard”
„ Current work being done so you don’t have to be a
wizard
‹Outline
„ TCP Overview
„ TCP Tuning Techniques (focus on Linux)
„ TCP Issues
„ Network Monitoring Tools
„ Current TCP Research
Brian L. Tierney
Slide: 3
How TCP works: A very short overview
‹ Congestion window (CWND) = the number of packets the
sender is allowed to send
„ The larger the window size, the higher the throughput
z Throughput = Window size / Round-trip Time
‹ TCP Slow start
„ exponentially increase the congestion window size until a packet is
lost
z this gets a rough estimate of the optimal congestion window
size
packet loss
timeout
CWND
slow start:
exponential
increase
congestion
avoidance:
linear
increase
retransmit:
slow start
again
time
Brian L. Tierney
Slide: 4
TCP Overview
‹ Congestion avoidance
„ additive increase: starting from the rough estimate, linearly
increase the congestion window size to probe for additional
available bandwidth
„ multiplicative decrease: cut congestion window size aggressively if
a timeout occurs
packet loss
timeout
CWND
slow start:
exponential
increase
congestion
avoidance:
linear
increase
retransmit:
slow start
again
time
Brian L. Tierney
Slide: 5
TCP Overview
‹ Fast Retransmit: retransmit after 3 duplicate acks (got 3 additional
packets without getting the one you are waiting for)
„ this prevents expensive timeouts
„ no need to go into “slow start” again
‹ At steady state, CWND oscillates around the optimal window size
‹ With a retransmission timeout, slow start is triggered again
packet loss
timeout
CWND
slow start:
exponential
increase
congestion
avoidance:
linear
increase
retransmit:
slow start
again
time
Brian L. Tierney
Slide: 6
Terminology
‹ The term “Network Throughput” is vague and should be
avoided
„ Capacity: link speed
z Narrow Link: link with the lowest capacity along a path
z Capacity of the end-to-end path = capacity of the narrow link
„ Utilized bandwidth: current traffic load
„ Available bandwidth: capacity – utilized bandwidth
z Tight Link: link with the least available bandwidth in a path
„ Achievable bandwidth: includes protocol and host issues
45 Mbps
10 Mbps
100 Mbps
45 Mbps
source
sink
Narrow Link
Brian L. Tierney
Tight Link
Slide: 7
More Terminology
‹RTT: Round-trip time
‹Bandwidth*Delay Product = BDP
„ The number of bytes in flight to fill the entire path
„ Example: 100 Mbps path; ping shows a 75 ms RTT
zBDP = 100 * 0.075 / 2 = 3.75 Mbits (470 KB)
‹LFN: Long Fat Networks
„ A network with a large BDP
Brian L. Tierney
Slide: 8
TCP Performance Tuning Issues
‹Getting good TCP performance over high-latency
high-bandwidth networks is not easy!
‹You must keep the pipe full, and the size of the
pipe is directly related to the network latency
„ Example: from LBNL (Berkeley, CA) to ANL (near
Chicago, IL), the narrow link is 1000 Mbits/sec, and the
one-way latency is 25ms
zBandwidth = 539 Mbits/sec (67 MBytes/sec) (OC12 =
622 Mbps - ATM and IP headers)
„ Need (1000 / 8) * .025 sec = 3.125 MBytes of data “in
flight” to fill the pipe
Brian L. Tierney
Slide: 9
Setting the TCP buffer sizes
‹ It is critical to use the optimal TCP send and receive
socket buffer sizes for the link you are using.
„ Recommended size = 2 x Bandwidth Delay Product (BDP)
„ if too small, the TCP window will never fully open up
„ if too large, the sender can overrun the receiver, and the TCP
window will shut down
‹ Default TCP buffer sizes are way too small for this type of
network
„ default TCP send/receive buffers are typically 64 KB
„ with default TCP buffers, you can only get a small % of the
available bandwidth!
Brian L. Tierney
Slide: 10
Throughput (Mbits/sec)
Importance of TCP Tuning
300
Tuned for
LAN
Tuned for
WAN
Tuned for
Both
264
264
200
152
112
100
112
44
64KB TCP
Buffers
512 KB TCP
Buffers
LAN (rtt = 1ms)
WAN (rtt = 50ms)
Brian L. Tierney
Slide: 11
TCP Buffer Tuning: System
‹ Need to adjust system max TCP buffer
„ Example: in Linux (2.4 and 2.6) add the entries below to the file
/etc/sysctl.conf, and then run "sysctl -p”
# increase TCP max buffer size
net.core.rmem_ max = 16777216
net.core.wmem_ max = 16777216
# increase Linux autotuning TCP buffer limits
# min, default, and max number of bytes to use
net.ipv4.tcp_rme m = 4096 87380 16777216
net.ipv4.tcp_w me m = 4096 65536 16777216
‹ Similar changes needed for other Unix OS’s
‹ For more info, see: http://dsd.lbl.gov/TCP-Tuning/
Brian L. Tierney
Slide: 12
TCP Buffer Tuning: Application
‹ Must adjust buffer size in your applications:
int skt, int sndsize = 2 * 1024 * 1024;
err = setsockopt(skt, SOL_SOCKET, SO_SNDBUF,
(char *)&sndsize,(int)sizeof(sndsize));
and/or
err = setsockopt(skt, SOL_SOCKET, SO_RCVBUF,
(char *)&sndsize,(int)sizeof(sndsize));
‹ It’s a good idea to check the following:
err = getsockopt(skt, SOL_S O C K ET, SO_RCVBUF,
(char *)&sockbufsize, &size);
If(size != sndsize)
printf(stderr,“ Warning: requested TCP buffer of %d, but only got %d
\n”, sndsize, size);
Brian L. Tierney
Slide: 13
Determining the Buffer Size
‹ The optimal buffer size is twice the bandwidth*delay
product of the link:
buffer size = 2 * bandwidth * delay
‹ The ping program can be used to get the delay
„ e.g.: >ping
-s 1500 lxplus.cern.ch
1500 bytes from lxplus012.cern.ch: icmp_seq=0. time=175. ms
1500 bytes from lxplus012.cern.ch: icmp_seq=1. time=176. ms
1500 bytes from lxplus012.cern.ch: icmp_seq=2. time=175. ms
‹ pipechar or pathrate can be used to get the bandwidth of
the slowest hop in your path. (see next slides)
‹ Since ping gives the round trip time (RTT), this formula
can be used instead of the previous one:
buffer size = bandwidth * RTT
Brian L. Tierney
Slide: 14
Buffer Size Example
‹ping time = 50 ms
‹Narrow link = 500 Mbps (62 Mbytes/sec)
„ e.g.: the end-to-end network consists of all 1000 BT
ethernet and OC-12 (622 Mbps)
‹TCP buffers should be:
„ .05 sec * 62 = 3.1 Mbytes
Brian L. Tierney
Slide: 15
Sample Buffer Sizes
‹ UK to...
„
„
„
„
UK (RTT = 5 ms, narrow link = 1000 Mbps) : 625 KB
Europe: (25 ms, narrow link = 500 Mbps): 1.56 MB
US: (150 ms, narrow link = 500 Mbps): 9.4 MB
Japan: (RTT = 260, narrow link = 150 Mbps): 4.9 MB
‹ Note: default buffer size is usually only 64 KB, and default
maximum buffer size for is only 256KB
„ Linux Autotuning default max = 128 KB;
‹ 10-150 times too small!
Brian L. Tierney
Slide: 16
More Problems: TCP congestion control
Path = LBL
to CERN
(Geneva)
OC-3, (in
2000), RTT
= 150 ms
average
BW =
30 Mbps
Brian L. Tierney
Slide: 17
Work-around: Use Parallel Streams
RTT = 70 ms
graph from Tom Dunigan, ORNL
Brian L. Tierney
Slide: 18
Throughput (Mbits/sec)
Tuned Buffers vs. Parallel Steams
30
25
20
15
10
5
0
no tuning
Brian L. Tierney
tuned
TCP
buffers
tuned
10
TCP
parallel
streams, buffers, 3
no tuning parallel
streams
Slide: 19
Parallel Streams Issues
‹Potentially unfair
‹Places more load
on the end hosts
‹But they are
necessary when
you don’t have root
access, and can’t
convince the
sysadmin to
increase the max
TCP buffers
graph from Tom Dunigan, ORNL
Brian L. Tierney
Slide: 20
NFNN2, 20th-21st June 2005
National e-Science Centre, Edinburgh
Network Monitoring Tools
http://gridmon.dl.ac.uk/nfnn/
traceroute
>traceroute pcgiga.cern.ch
traceroute to pcgiga.cern.ch (192.91.245.29), 30 hops max, 40 byte packets
1 ir100gw-r2.lbl.gov (131.243.2.1) 0.49 ms
0.26 ms
0.23 ms
2 er100gw.lbl.gov (131.243.128.5) 0.68 ms
0.54 ms
0.54 ms
3 198.129.224.5 (198.129.224.5) 1.00 ms *d9* 1.29 ms
4 lbl2-ge-lbnl.es.net (198.129.224.2) 0.47 ms
0.59 ms
0.53 ms
5 snv-lbl-oc48.es.net (134.55.209.5) 57.88 ms
56.62 ms
61.33 ms
6 chi-s-snv.es.net (134.55.205.102) 50.57 ms
49.96 ms
49.84 ms
7 ar1-chicago-esnet.cern.ch (198.124.216.73) 50.74 ms 51.15 ms 50.96 ms
8 cernh9-pos100.cern.ch (192.65.184.34) 175.63 ms
176.05 ms
176.05 ms
9 cernh4.cern.ch (192.65.185.4) 175.92 ms
175.72 ms
176.09 ms
10 pcgiga.cern.ch (192.91.245.29)
175.58 ms
175.44 ms
175.96 ms
Can often learn about the network from the router names:
ge = Gigabit Ethernet
oc48 = 2.4 Gbps (oc3 = 155 Mbps, oc12=622 Mbps)
Brian L. Tierney
Slide: 22
Iperf
‹ iperf : very nice tool for measuring end-to-end TCP/UDP performance
„ http://dast.nlanr.net/Projects/Iperf/
„ Can be quite intrusive to the network
‹ Example:
„ Server: iperf -s -w 2M
„ Client: iperf -c hostname -i 2 -t 20 -l 128K -w 2M
Client connecting to hostname
[ ID] Interval
Transfer Bandwidth
[ 3] 0.0- 2.0 sec 66.0 MBytes 275 Mbits/sec
[ 3] 2.0- 4.0 sec 107 MBytes 451 Mbits/sec
[ 3] 4.0- 6.0 sec 106 MBytes 446 Mbits/sec
[ 3] 6.0- 8.0 sec 107 MBytes 443 Mbits/sec
[ 3] 8.0-10.0 sec 106 MBytes 447 Mbits/sec
[ 3] 10.0-12.0 sec 106 MBytes 446 Mbits/sec
[ 3] 12.0-14.0 sec 107 MBytes 450 Mbits/sec
[ 3] 14.0-16.0 sec 106 MBytes 445 Mbits/sec
[ 3] 16.0-24.3 sec 58.8 MBytes 59.1 Mbits/sec
[ 3] 0.0-24.6 sec 871 MBytes 297 Mbits/sec
Brian L. Tierney
Slide: 23
pathrate / pathload
‹Nice tools from Georgia Tech:
„ pathrate: measures the capacity of the narrow link
„ pathload: measures the available bandwidth
‹Both work pretty well.
„ pathrate can take a long time (up to 20 minutes)
„ These tools attempt to be non-intrusive
‹Open Source; available from:
„ http://www.pathrate.org/
Brian L. Tierney
Slide: 24
pipechar
‹ Tool to measure hop-by-hop available bandwidth,
capacity, and congestion
‹ Takes 1-2 minutes to measure an 8 hop path
‹ But not always accurate
„ Results affected by host speed
z Hard to measure links faster than host interface
„ Results after a slow hop typically not accurate, for example, if the
first hop is a wireless link, and all other hops are 100 BT or faster,
then results are not accurate
‹ client-side only tool: puts very little load on the network
(about 100 Kbits/sec)
‹ Available from: http://dsd.lbl.gov/NCS/
„ part of the netest package
Brian L. Tierney
Slide: 25
pipechar output
dpsslx04.lbl.gov(59)>pipechar firebird.ccs.ornl.gov
PipeChar statistics: 82.61% reliable
Fro m localhost:
827.586 Mbps GigE (1020.4638 M bps)
1: ir100gw-r2.lbl.gov
(131.243.2.1 )
|
1038.492 Mbps GigE
<11.2000 % BW used>
2: er100gw.lbl.gov
(131.243.128.5)
|
1039.246 Mbps GigE
<11.2000 % BW used>
3: lbl2-ge-lbnl.es.net
(198.129.224.2)
|
285.646 Mbps congested bottleneck <71.2000% B W used>
4: snv-lbl-oc48.es.net
(134.55.209.5)
|
9935.817 Mbps OC192 <94.0002 % BW used>
5: orn-s-snv.es.net
(134.55.205.121)
|
341.998 Mbps congested bottleneck <65.2175 % B W used>
6: ornl-orn.es.net
(134.55.208.62)
|
298.089 Mbps congested bottleneck <70.0007 % B W used>
7: orgwy-ext.ornl.gov
(192.31.96.225)
339.623 Mbps congested bottleneck <65.5502 % B W used>
|
8: ornlgwy-ext.ens.ornl.gov
(198.124.42.162)
|
232.005 Mbps congested bottleneck <76.6233 % B W used>
9: ccsrtr.ccs.ornl.gov
(160.91.0.66 )
|
268.651 Mbps GigE (1023.4655 Mbps)
10: firebird.ccs.ornl.gov
(160.91.192.165)
Brian L. Tierney
Slide: 26
tcpdump / tcptrace
‹ tcpdump: dump all TCP header information for a specified
source/destination
„ ftp://ftp.ee.lbl.gov/
‹ tcptrace: format tcpdump output for analysis using xplot
„ http://www.tcptrace.org/
„ NLANR TCP Testrig : Nice wrapper for tcpdump and tcptrace tools
z http://www.ncne.nlanr.net/TCP/testrig/
‹ Sample use:
tcpdump -s 100 -w /tmp/tcpdump.out host hostname
tcptrace -Sl /tmp/tcpdump.out
xplot /tmp/a2b_tsg.xpl
Brian L. Tierney
Slide: 27
tcptrace and xplot
‹X axis is time
‹Y axis is sequence number
‹the slope of this curve gives the throughput over
time.
‹xplot tool make it easy to zoom in
Brian L. Tierney
Slide: 28
Zoomed In View
‹
‹
‹
‹
Green Line: ACK values received from the receiver
Yellow Line tracks the receive window advertised from the receiver
Green Ticks track the duplicate ACKs received.
Yellow Ticks track the window advertisements that were the same as
the last advertisement.
‹ White Arrows represent segments sent.
‹ Red Arrows (R) represent retransmitted segments
Brian L. Tierney
Slide: 29
Other Tools
‹NLANR Tools Repository:
„ http://www.ncne.nlanr.net/software/tools/
‹SLAC Network MonitoringTools List:
„ http://www.slac.stanford.edu/xorg/nmtf/nmtf-tools.html
Brian L. Tierney
Slide: 30
Other TCP Issues
‹Things to be aware of:
„ TCP slow-start
zOn a path with a 50 ms RTT, it takes 12 RTT’s to
ramp up to full window size, so need to send about
10 MB of data before the TCP congestion window
will fully open up.
„ host issues
zMemory copy speed
zI/O Bus speed
zDisk speed
Brian L. Tierney
Slide: 31
TCP Slow Start
Brian L. Tierney
Slide: 32
Duplex Mismatch Issues
‹A common source of trouble with Ethernet
networks is that the host is set to full duplex, but
the Ethernet switch is set to half-duplex, or visa
versa.
‹Most newer hardware will auto-negotiate this, but
with some older hardware, auto-negotiation
sometimes fails
„ result is a working but very slow network (typically only
1-2 Mbps)
„ best for both to be in full duplex if possible, but some
older 100BT equipment only supports half-duplex
‹NDT is a good tool for finding duplex issues:
„ http://e2epi.internet2.edu/ndt/
Brian L. Tierney
Slide: 33
Jumbo Frames
‹ Standard Ethernet packet is 1500 bytes (aka: MTU)
‹ Some gigabit Ethernet hardware supports “jumbo frames”
(jumbo packet) up to 9 KBytes
„ This helps performance by reducing the number of host interrupts
„ Some jumbo frame implementations do not interoperate
„ Most routers allow at most 4K MTUs
‹ First Ethernet was 3 Mbps (1972)
‹ First 10 Gbit/sec Ethernet hardware: 2001
„ Ethernet speeds have increased 3000x since the 1500 byte frame
was defined
„ Computers now have to work 3000x harder to keep the network full
Brian L. Tierney
Slide: 34
Linux Autotuning
‹ Sender-side TCP buffer autotuning introduced in Linux 2.4
„ TCP send buffer starts at 64 KB
„ As the data transfer takes place, the buffer size is continuously readjusted up max autotune size (default = 128K)
‹ Need to increase defaults: (in /etc/sysctl.conf)
# increase TCP max buffer size
net.core.rme m _max = 16777216
net.core.wmem_ max = 16777216
# increase Linux autotuning TCP bufferlimits
# min, default, and max number of bytes to use
net.ipv4.tcp_rme m = 4096 87380 16777216
net.ipv4.tcp_w me m = 4096 65536 16777216
‹ Receive buffers need to be bigger than largest send buffer used
„ Use setsockopt() call
Brian L. Tierney
Slide: 35
Linux 2.4 Issues
‹ ssthresh caching
„ ssthresh (Slow Start Threshold): size of CWND to use when
switching from exponential increase to linear increase
„ The value for ssthresh for a given path is cached in the routing
table.
„ If there is a retransmission on a connection to a given host, then all
connections to that host for the next 10 minutes will use a reduced
ssthresh.
„ Or, if the previous connect to that host is particularly good, then
you might stay in slow start longer, so it depends on the path
„ The only way to disable this behavior is to do the following before
all new connections (you must be root):
z sysctl -w net.ipv4.route.flush=1
„ The web100 kernel patch adds a mechanism to permanently
disable this behavior:
z sysctl -w net.ipv4.web100_no_metrics_save = 1
Brian L. Tierney
Slide: 36
ssthresh caching
•The value of
CWND where
this loss
happened will
get cached
Brian L. Tierney
Slide: 37
Linux 2.4 Issues (cont.)
‹ SACK implementation problem
„ For very large BDP paths where the TCP window is > 20 MB, you
are likely to hit the Linux SACK implementation problem.
„ If Linux has too many packets in flight when it gets a SACK event, it
takes too long to located the SACKed packet,
z you get a TCP timeout and CWND goes back to 1 packet.
„ Restricting the TCP buffer size to about 12 MB seems to avoid this
problem, but limits your throughput.
„ Another solution is to disable SACK.
sysctl-w net.ipv4.tcp_sack = 0
„ This is still a problem in 2.6, but they are working on a solution
‹ Transmit queue overflow
„ If the interface transmit queue overflows, the Linux TCP stack
treats this as a retransmission.
„ Increasing txqueuelen can help:
ifconfig eth0 txqueuelen 1000
Brian L. Tierney
Slide: 38
NFNN2, 20th-21st June 2005
National e-Science Centre, Edinburgh
Recent/Current TCP Work
http://gridmon.dl.ac.uk/nfnn/
TCP Response Function
‹Well known fact that TCP does not scale to highspeed networks
‹Average TCP congestion window = 1.2 p
segments
„ p = packet loss rate
‹What this means:
„ For a TCP connection with 1500-byte packets and a
100 ms round-trip time, filling a 10 Gbps pipe would
require a congestion window of 83,333 packets, and a
packet drop rate of at most one drop every
5,000,000,000 packets.
„ requires at most one packet loss every 6000s, or
1h:40m to keep the pipe full
Brian L. Tierney
Slide: 40
Proposed TCP Modifications
‹High Speed TCP: Sally Floyd
„ http://www.icir.org/floyd/hstcp.html
‹BIC/CUBIC:
„ http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/
‹LTCP (Layered TCP)
„ http://students.cs.tamu.edu/sumitha/research.html
‹HTCP: (Hamilton TCP)
„ http://www.hamilton.ie/net//htcp/
‹Scalable TCP
„ http://www-lce.eng.cam.ac.uk/~ctk21/scalable/
Brian L. Tierney
Slide: 41
Proposed TCP Modifications (cont.)
‹ XCP:
„ XCP rapidly converges on the optimal congestion window using a
completely new router paradigm.
z This makes it very difficult to deploy and test
„ http://www.ana.lcs.mit.edu/dina/XCP/
‹ FAST TCP:
„ http://netlab.caltech.edu/FAST/
‹ Each if these alternatives give roughly similar throughput
„ Vary mainly in “stability” and “friendliness” with other protocols
‹ Each of these require sender-side only modifications to
standard TCP
Brian L. Tierney
Slide: 42
TCP: Reno vs. BIC
‹TCP-Reno
(Linux 2.4)
‹BIC-TCP
(Linux 2.6)
Brian L. Tierney
Slide: 43
TCP: Reno vs. BIC
• BIC-TCP
recovers from
loss more
aggressively
than TCP-Reno
Brian L. Tierney
Slide: 44
Sample Results
From Doug Leith, Hamilton Institute, http://www.hamilton.ie/net/eval/
Fairness
Between Flows
Link Utilization
Brian L. Tierney
Slide: 45
New Linux 2.6 changes
‹ Added receive buffer autotuning: adjust receive window
based on RTT
„ sysctlnet.ipv4.tcp_moderate_rcvbuf
„ Still need to increase max value: net.ipv4.tcp_rme m
‹ Starting in Linux 2.6.7 (and back-ported to 2.4.27), BIC
TCP is part of the kernel, and enabled by default.
‹ Bug found that caused performance problems under some
circumstances, fixed in 2.6.11.
‹ Added ability to disable ssthresh caching (like web100)
net.ipv4.tcp_no_metrics_save = 1
Brian L. Tierney
Slide: 46
Linux 2.6 Issues
‹"tcp segmentation offload” issue:
„ Linux 2.6 ( < 2.6.11) has bug with certain Gigabit and
10 Gig ethernet drivers and NICs that support "tcp
segmentation offload",
z These include Intel e1000 and ixgb drivers,
Broadcom tg3, and the s2io 10 GigE drivers.
zTo fix this problem, use ethtool to disable
segmentation offload:
ethtool -K eth0 tso off
„ Bug fixed in Linux 2.6.12
Brian L. Tierney
Slide: 47
Linux 2.6.12-rc3 Results
Path
Linux 2.4
LBL to ORNL
300 Mbps
Linux 2.6 with Linux 2.6, no
BIC
BIC
700 Mbps
500 Mbps
300 Mbps
830 Mbps
625 Mbps
70 Mbps
560 Mbps
140 Mbps
RTT = 67 ms
LBL to PSC
RTT = 83 ms
LBL to IE
RTT = 153 ms
Results = Peak Speed during 3 minute test
Note: BIC is ON by default in Linux 2.6
Sending host = 2.8 GHz Intel Xeon with Intel e1000 NIC
Brian L. Tierney
Slide: 48
Linux 2.6.12-rc3 Results
TCP Results
800
700
600
500
Linux 2.6, BIC TCP
400
Linux 2.4
Linux 2.6, BIC off
300
200
RTT = 67 ms
100
0
1
3
5
7
9 11 13 15 17 19 21 23 25 27 29 31 33 35
time slot (5 second intervals)
Brian L. Tierney
Slide: 49
Remaining Linux BIC Issues
‹ But: on
some paths
BIC still
seems to
have
problems…
RTT = 83 ms
Brian L. Tierney
Slide: 50
NFNN2, 20th-21st June 2005
National e-Science Centre, Edinburgh
Application Performance Issues
http://gridmon.dl.ac.uk/nfnn/
Techniques to Achieve High
Throughput over a WAN
‹Consider using multiple TCP sockets for the data
stream
‹Use a separate thread for each socket
‹Keep the data pipeline full
„ use asynchronous I/O
zoverlap I/O and computation
„ read and write large amounts of data (> 1MB) at a time
whenever possible
„ pre-fetch data whenever possible
‹Avoid unnecessary data copies
„ manipulate pointers to data blocks instead
Brian L. Tierney
Slide: 52
Use Asynchronous I/O
‹ I/O followed by
processing
Next IO starts
when processing
ends
‹ overlapped I/O and
processing
process previous
block
almost a 2:1 speedup
Brian L. Tierney
remote IO
Slide: 53
Throughput vs. Latency
‹ Most of the techniques we have discussed are designed to
improve throughput
‹ Some of these might increase latency
„ with large TCP buffers, OS will buffer more data before sending it
‹ Goal of a distributed application programmer:
„ hide latency
‹ Some techniques to help decrease latency:
„ use separate control and data sockets
„ use TCP_NODELAY option on control socket
z combine control messages together into 1 larger message
whenever possible on TCP_NODELAY sockets
Brian L. Tierney
Slide: 54
scp Issues
‹Don’t use scp to copy large files!
„ scp has its own internal buffering/windowing that
prevents it from ever being able to fill LFNs!
‹Explanation of problem and openssh patch
solution from PSC
„ http://www.psc.edu/networking/projects/hpn-ssh/
Brian L. Tierney
Slide: 55
Conclusions
‹ The wizard gap is starting to close (slowly)
„ If max TCP buffers are increased
‹ Tuning TCP is not easy!
„ no single solution fits all situations
z need to be careful TCP buffers are not too big or too small
z sometimes parallel streams help throughput, sometimes they hurt
„ Linux 2.6 helps a lot
‹ Design your network application to be as flexible as possible
„ make it easy for clients/users to set the TCP buffer sizes
„ make it possible to turn on/off parallel socket transfers
z probably off by default
‹ Design your application for the future
„ even if your current WAN connection is only 45 Mbps (or less), some day it
will be much higher, and these issues will become even more important
Brian L. Tierney
Slide: 56
For More Information
http://dsd.lbl.gov/TCP-tuning/
„ links to all network tools mentioned here
„ sample TCP buffer tuning code, etc.
BLTierney@LBL.GOV
Brian L. Tierney
Slide: 57